Re: BTF loading failing on perf

[Arnaldo Carvalho de Melo]

On Fri, Jun 06, 2025 at 09:20:57AM -0700, Andrii Nakryiko wrote:
> On Fri, Jun 6, 2025 at 9:14 AM Leo Yan <leo.yan@xxxxxxx> wrote:
> > On Fri, Jun 06, 2025 at 12:37:55PM -0300, Arnaldo Carvalho de Melo wrote:
> > > root@number:~# perf trace -e openat --max-events=1
> > > libbpf: failed to read kernel BTF from '/sys/kernel/btf/vmlinux': -ENODEV
> > > libbpf: failed to read kernel BTF from '/sys/kernel/btf/vmlinux': -ENODEV
> > >      0.000 ( 0.016 ms): ptyxis-agent/4375 openat(dfd: CWD, filename: "/proc/6593/cmdline", flags: RDONLY|CLOEXEC) = 13
> > > root@number:~#
> > >
> > > openat(AT_FDCWD, "/sys/kernel/btf/vmlinux", O_RDONLY) = 258
> > > mmap(NULL, 6519699, PROT_READ, MAP_PRIVATE, 258, 0) = -1 ENODEV (No such device)
> > > libbpf: failed to read kernel BTF from '/sys/kernel/btf/vmlinux': -ENODEV
> >
> > Have you included the commit below in the kernel side?
> 
> It doesn't matter, libbpf should silently fallback to non-mmap() way,

Right, it has to work with older kernels, etc.

> and it clearly doesn't.

> We need something like this:
> +++ b/tools/lib/bpf/btf.c
> @@ -1384,12 +1384,12 @@ static struct btf *btf_parse_raw_mmap(const
> char *path, struct btf *base_btf)
> 
>         fd = open(path, O_RDONLY);
>         if (fd < 0)
> -               return libbpf_err_ptr(-errno);
> +               return ERR_PTR(-errno);
> 
>         if (fstat(fd, &st) < 0) {
>                 err = -errno;
>                 close(fd);
> -               return libbpf_err_ptr(err);
> +               return ERR_PTR(err);
>         }
> 
>         data = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
> @@ -1397,7 +1397,7 @@ static struct btf *btf_parse_raw_mmap(const char
> *path, struct btf *base_btf)
>         close(fd);
> 
>         if (data == MAP_FAILED)
> -               return libbpf_err_ptr(err);
> +               return ERR_PTR(err);
> 
>         btf = btf_new(data, st.st_size, base_btf, true);
>         if (IS_ERR(btf))
> 
> libbpf_err_ptr() should be used for user-facing API functions, they
> return NULL on error and set errno, so checking for IS_ERR() is wrong
> here.

And the only user of the above function is:

                btf = btf_parse_raw_mmap(sysfs_btf_path, NULL);
                if (IS_ERR(btf))
                        btf = btf__parse(sysfs_btf_path, NULL);

That expects ERR_PTR() to then use IS_ERR().

I think this could be automated with something like coccinnele(sp)?

Anyway, I tested the patch above and it seems to fix the issue, so:

Reported-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Reviewed-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Tested-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>

- Arnaldo

 
> Lorenz, can you please test and send a proper fix ASAP?
> 
> >
> > commit a539e2a6d51d1c12d89eec149ccc72ec561639bc
> > Author: Lorenz Bauer <lmb@xxxxxxxxxxxxx>
> > Date:   Tue May 20 14:01:17 2025 +0100
> >
> >     btf: Allow mmap of vmlinux btf
> >
> >     User space needs access to kernel BTF for many modern features of BPF.
> >     Right now each process needs to read the BTF blob either in pieces or
> >     as a whole. Allow mmaping the sysfs file so that processes can directly
> >     access the memory allocated for it in the kernel.
> >
> >     remap_pfn_range is used instead of vm_insert_page due to aarch64
> >     compatibility issues.
> >
> >     Signed-off-by: Lorenz Bauer <lmb@xxxxxxxxxxxxx>
> >     Signed-off-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
> >     Tested-by: Alan Maguire <alan.maguire@xxxxxxxxxx>
> >     Reviewed-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> >     Link: https://lore.kernel.org/bpf/20250520-vmlinux-mmap-v5-1-e8c941acc414@xxxxxxxxxxxxx
> >
> > Thanks,
> > Leo